Artificial intelligence in healthcare has been a recurring topic of discussion for years. A breakthrough would be announced by someone. A tool would be tested in a hospital. The next step would be a press release full of words about revolution and change. The pilot would then end quietly. The instrument would remain idle. They would all wait for the next announcement.
It appears that the loop is breaking. Instead of making a drastic change, the excitement is beginning to give way to something slower and, to be honest, more intriguing: a creeping seriousness.
Approximately 24,000 leaders in the healthcare and technology sectors convened at HIMSS26 in Las Vegas this past March in a large, carpeted conference room that makes everything seem equally significant and equally unimportant. However, this time felt different. The discussions did not center on the potential applications of AI. They discussed what happens when a system behaves poorly under clinical pressure, what accountability looks like when a system makes a mistake, and whether any of this truly reduces costs or merely shifts them. It’s important to focus on that transition from possibility to accountability.
Bob Wachter, the chair of UCSF’s Department of Medicine, has been observing this field long enough to see the shift. He spent years documenting how electronic health records subtly made doctors miserable. He acknowledged that his 2015 book was grumpy. He had witnessed hospitals spend billions on digital systems that transformed physicians into data-entry clerks, obliterating the silent cornerstone of medicine—eye contact. Something changed when he used ChatGPT for the first time in late 2022.
It wasn’t because the technology was flawless, but rather because it was the first tool that truly allowed clinicians to interact with data in an understandable manner and obtain valuable information. Since then, he has been paying attention, and his most recent writing exhibits a cautious but sincere enthusiasm—the kind that results from seeing something fail for a long time before it gradually begins to work.
One of the more underappreciated trends in tech history is the productivity gap Wachter describes, where general-purpose technologies fall short for ten years before delivering. With electrification, it took place. With the internet, it took place. Though it’s still too early to say with certainty, it might be happening right now with AI in clinical settings.
It is evident that the questions being posed have evolved. Executives at the conference pushed vendors not only on capability but also on failure modes, how their systems would function under regulatory pressure, and what governance structures actually looked like in real-world scenarios. When Dr. Niki Panich of Penguin AI stated that governance was now being viewed as a design requirement rather than a compliance checkbox, she captured something genuine. That’s an important distinction. It indicates that before the product was shipped, someone, somewhere, began to consider trust.
There is still a lot of skepticism to cling to. Large language models by themselves don’t seem to be the solution, and operating them at scale within health systems comes at a high cost. For many organizations, the return on investment is still unclear. Furthermore, despite the polished appearance of the conference booths, there is still worry that AI will reinforce preexisting biases or subtly increase harm, particularly for patients in underserved or rural areas.
However, it’s difficult to ignore the fact that the discussion has shifted from whether AI should be used in healthcare to how to make it responsible when it already exists. That’s a big change. It indicates that the technology has reached irreversibility rather than perfection. Convincing health systems to implement AI is currently a challenge. It is ensuring that adoption does not take precedence over judgment.
That is a more difficult issue. It also seems to be the more significant one.
